hparams.py

def create_hparams():
    """Create model hyperparameters. Parse nondefault from given string."""

    hparams = dict(
        ################################
        # Experiment Parameters        #
        ################################
        epochs=500,
        iters_per_checkpoint=1000,
        seed=1234,
        dynamic_loss_scaling=True,
        fp16_run=False,
        distributed_run=False,
        dist_backend="nccl",
        dist_url="tcp://localhost:54321",
        cudnn_enabled=True,
        cudnn_benchmark=False,
        ignore_layers=['embedding.weight'],

        ################################
        # Data Parameters             #
        ################################
        load_mel_from_disk=False,
        training_files='filelists/ljs_audio_text_train_filelist.txt',
        validation_files='filelists/ljs_audio_text_val_filelist.txt',
        text_cleaners=['english_cleaners'],

        ################################
        # Audio Parameters             #
        ################################
        # max_wav_value=32768.0,
        sampling_rate=16000, #22050,
        filter_length=1024,
        hop_length=256,
        win_length=1024,
        n_mel_channels=80,
        mel_fmin=0.0,
        mel_fmax=8000.0,

        ################################
        # Model Parameters             #
        ################################
        # n_symbols=len(symbols),
        # symbols_embedding_dim=512,

    
        # Encoder parameters
        num_init_filters= 24,
        encoder_kernel_size = 5,
        encoder_n_convolutions=5,
        encoder_embedding_dim=1024,

        # Decoder parameters
        n_frames_per_step=1,  # currently only 1 is supported
        decoder_rnn_dim=1024,
        prenet_dim=256,
        max_decoder_steps=300,
        gate_threshold=0.5,
        p_attention_dropout=0.1,
        p_decoder_dropout=0.1,

        # Attention parameters
        attention_rnn_dim=1024,
        attention_dim=128,

        # Location Layer parameters
        attention_location_n_filters=32,
        attention_location_kernel_size=31,

        # Mel-post processing network parameters
        postnet_embedding_dim=512,
        postnet_kernel_size=5,
        postnet_n_convolutions=5,

        ################################
        # Optimization Hyperparameters #
        ################################
        use_saved_learning_rate=False,
        learning_rate=1e-4,
        weight_decay=1e-6,
        grad_clip_thresh=1.0,
        batch_size=64,
        mask_padding=True,  # set model's padded outputs to padded values

        teacher_forcing_probability=0.5,
    )
    
    class HParams:
        def __init__(self, dictionary):
            for k, v in dictionary.items():
                setattr(self, k, v)

    hparams = HParams(hparams)
    
    # if hparams_string:
    #     tf.logging.info('Parsing command line hparams: %s', hparams_string)
    #     hparams.parse(hparams_string)

    # if verbose:
    #     tf.logging.info('Final parsed hparams: %s', hparams.values())

    return hparams